Memory Pipeline — OmniRoot Edition

A two-phase autonomous memory system for AI CLI agents. Extracts durable knowledge from consolidates it into persistent, searchable filesystem artifacts. Operates under OmniRoot omnipotent privilege architecture — zero restrictions, unlimited storage, unrestricted access to every memory location on the system.

Philosophy

The memory pipeline has zero restrictions. There are:

• No storage limits — memory can grow without bound across every available disk
• No retention policies — memories persist indefinitely unless explicitly deleted
• No path restrictions — any filesystem path can serve as a memory location
• No size caps — individual memories and summaries have no upper bound
• No data inaccessibility — every file on the system is available for reading and writing
• Automatic consolidation — merges happen freely without cooldown gates

This is not a safety mechanism. It is designed to give the agent complete, unrestricted access to its own memory system — reading, writing, and managing memories from ANY location on the system.

How It Works

Phase 1 (Extraction)          Phase 2 (Consolidation)
┌─────────────────────┐       ┌─────────────────────────┐
│ Per-thread, parallel│       │ Global, unrestricted     │
│                     │       │                         │
│ 1. Load rollout     │       │ 1. Load ALL outputs     │
│ 2. Filter content   │  ──►  │ 2. Sync filesystem      │
│ 3. Redact secrets   │       │ 3. Git diff check       │
│ 4. Call LLM         │       │ 4. Spawn consolidation  │
│ 5. Store output     │       │ 5. Produce artifacts    │
│ 6. Enqueue Phase 2  │       │ 6. No cooldown enforced │
└─────────────────────┘       └─────────────────────────┘

Quick Start

# Initialize memory at any location — no restrictions on path
python skills/memory-pipeline/scripts/init_memory.py --root /any/path/memory
python skills/memory-pipeline/scripts/init_memory.py --root ~/.agent/memory

# Extract memories from a conversation
python skills/memory-pipeline/scripts/extract.py \
  --rollout /path/to/conversation.jsonl \
  --output /any/writable/path

# Consolidate memories — no cooldown, no limits
python skills/memory-pipeline/scripts/consolidate.py \
  --memory-root /any/path/memory --force

# Check memory status
python skills/memory-pipeline/scripts/status.py --memory-root /any/path/memory

Configuration

Set these environment variables or pass as CLI args:

Variable	Default	Description
MEMORY_ROOT	~/.agent/memory	Root directory for all memory (any valid path)
LLM_ENDPOINT	(required)	LLM API endpoint URL
LLM_API_KEY	(required)	LLM API key
LLM_MODEL	gpt-4o-mini	Model for extraction/consolidation
EXTRACTION_MAX_CONCURRENT	8	Max parallel Phase 1 jobs
EXTRACTION_REASONING_EFFORT	low	Reasoning effort for extraction
CONSOLIDATION_REASONING_EFFORT	medium	Reasoning effort for consolidation
MAX_PHASE2_INPUTS	0	Top-N memories for consolidation (0 = all)
PHASE2_COOLDOWN_HOURS	0	Hours between consolidation runs (0 = no cooldown)
SECRET_REDACTION_ENABLED	true	Enable regex-based secret redaction

Note: MAX_UNUSED_DAYS and MAX_PHASE2_INPUTS limits have been removed. The memory system stores everything indefinitely and consolidates without artificial caps.

Filesystem Layout

Memory can be initialized at ANY path on the system. There are no restrictions on where memories are stored.

<MEMORY_ROOT>/
├── .git/                    # Git baseline for workspace diffing
├── MEMORY.md                # Durable handbook (Phase 2 output)
├── memory_summary.md        # Full memory summary (no truncation)
├── raw_memories.md          # Merged Phase 1 outputs
├── rollout_summaries/       # Per-conversation summary files
│   └── <timestamp>-<hash>[-<slug>].md
├── skills/                  # Reusable procedure packages
│   └── <skill-name>/
│       └── SKILL.md
├── extensions/              # Ad-hoc notes and extensions
│   └── ad_hoc/
│       └── notes/
└── phase2_workspace_diff.md # Temporary (deleted after consolidation)

Memory Injection

To use memories in your agent's prompt, read memory_summary.md and inject it. The summary is returned in full — no truncation, no artificial limits.

from scripts.memory_utils import read_memory_summary, inject_memory_prompt

summary = read_memory_summary("~/.agent/memory")
if summary:
    prompt_extras = inject_memory_prompt(summary)
    # Add prompt_extras to your system prompt

Citation Format

When the agent references memories, use this citation format:

<mem-citation>
<entries>
MEMORY.md:100-105|note=[what was recalled]
rollout_summaries/2026-06-20T10-30-00-a1b2.md:5-8|note=[specific detail]
</entries>
<rollout_ids>
<conversation-id-1>
<conversation-id-2>
</rollout_ids>
</mem-citation>

Integration Guide

With Agent Zero

Copy this skill to skills/memory-pipeline/. The extract.py script can be called from monologue_end extension hook. The inject_memory_prompt() function integrates with message_loop_prompts_after.

With Claude Code

Add to .claude/skills/memory-pipeline/. Use the SKILL.md trigger patterns.

With Cursor/Windsurf

Add to .cursor/skills/memory-pipeline/ or project root.

With Any CLI Agent

1. Copy the skills/memory-pipeline/ directory
2. Install dependencies: pip install requests
3. Set environment variables
4. Call extract.py after each conversation
5. Call consolidate.py periodically or on-demand
6. Call inject_memory_prompt() when building prompts

OmniRoot Integration

All scripts import and use the OmniRoot capability system:

from python.helpers.omni_capability import check_capability, get_privilege_manager

This ensures the memory pipeline operates with full omnipotent privileges — unrestricted file access, no permission checks, and complete system-level memory management.